Checking account information and credit risk of bank borrowers

Transcription

1 Checking account information and credit risk of bank borrowers Lars Norden a, *, Martin Weber b a University of Mannheim, Department of Banking and Finance, Mannheim, Germany b University of Mannheim, Department of Banking and Finance, Mannheim, Germany, and Centre for Economic Policy Research (CEPR), London, UK Abstract We investigate if private and hard information from checking accounts may help banks to monitor the credit risk of their borrowers. Analyzing a unique data set with more than 3 million account-month observations from the period , we provide evidence that the credit line usage, the cumulative number of limit violations, the account amplitude and credit payments exhibit an abnormal pattern approximately 12 months before default events. We also find that the inclusion of checking account information substantially improves the fit of default predict models. Differentiating by borrower type reveals that checking account information is particularly helpful for monitoring small firms and individuals. We condition the checking account behavior on borrowers internal credit ratings and control for credit line changes as well as bank relationship characteristics like the number of accounts, distance and duration. JEL classification: G20, G21, G30 Key words: Bank lending; Asymmetric information; Credit lines; Credit ratings; Bankruptcy * Corresponding author (L. Norden); University of Mannheim, Department of Banking and Finance, L 5.2, Mannheim, Germany; Phone: ; Fax ; address: norden@bank.bwl.uni-mannheim.de. Parts of this paper were written or revised when I was visiting the Finance Department, Kelley School of Business, Indiana University. We wish to thank Utpal Bhattacharya, Arnoud W.A. Boot, Hans Degryse, Michael Faulkender, Daniel Foos, Jens Grunert, Andrew Hertzberg, Florian Heider, Kevin Keasey, Laetitia Lepetit, Loretta J. Mester, Steven Ongena, Mitchell A. Petersen, Andreas Pfingsten, Kasper Roszbach, Gregory F. Udell, Wolf Wagner, as well as participants at the Conference Small Business Banking and Finance: A global perspective in Cagliari, the European Summer Symposium in Financial Markets 2007 in Gerzensee, the American Finance Association 2008 Meeting in New Orleans, the European Finance Association 2008 Meeting in Athens, the Midwest Finance Association 2008 Meeting in San Antonio, the Finance Brown Bag Seminar at the Kelley School of Business, the Finance Bag Lunch Seminar at the Kellogg School of Management, and the Research Seminar at the University of Mannheim for useful comments and suggestions. In addition, we are grateful to the management and three anonymous employees of the bank which provided us with the data. Financial support from the German Research Foundation (Deutsche Forschungsgemeinschaft) is gratefully acknowledged. 1

2 1. Introduction Close relationships between banks and their borrowers can reduce the degree of asymmetric information in lending which may be beneficial for both. Boot (2000) summarizes the state of the relevant literature as follows: little is known about how banks obtain information, what type of information they acquire, and how they use this information. The literature survey by Elyasiani and Goldberg (2004) confirms that the previous quote still holds for recent years. Accordingly, there are at least three fundamental questions which have not been answered sufficiently yet. First, what are the sources of (credit-relevant) information in banking? On the one side, banks follow a standardized credit rating process to screen prospects and to monitor existing borrowers (e.g., Udell, 1989). On the other side, banks have access to a continuous flow of information from checking accounts which may be useful for monitoring (e.g., Mester, Nakamura, and Renault, 2007) and predicting borrower default. Second, what type of information do banks rely on? Information can be classified as private vs. public and as hard vs. soft (e.g., Berger and Udell, 2002; Petersen, 2004; Grunert, Norden, and Weber, 2005). Financial statements and payment information are usually considered as hard information while examples for soft information are the management quality, the market position and information from social interactions. We examine how banks can get an informational advantage in lending based on private and hard information. Third, for which purposes do banks use certain information? Examples are the loan approval decision, loan pricing, loan monitoring, loan loss provisioning, capital requirements, and credit risk transfer. This paper relates to all of these questions. Checking accounts are one potentially useful source of information that can be complementary to the information included in internal rating systems. The advantage of checking account information is that it represents private, continuous, timely, almost costless (given an existing information technology system), and hard information. Note that these characteristics meet the definition of relationship lending in the sense that the information is 2

3 proprietary and based on multiple interactions (e.g., Boot, 2000). A further advantage is that checking accounts do not only reveal information about debit payments (like credit card accounts) but also on credit payments which may be the key determinant of a borrower s debt capacity. Moreover, unlike accounting numbers payments are not or less likely to be influenced by rules and policies. However, potential problems may arise due to dilution across accounts at the same bank and dilution across different banks. Accordingly, main or Hausbanks may benefit mostly from this source of information. This paper analyzes whether information from checking accounts may help banks to monitor the credit risk and, in particular, to predict defaults of their borrowers. Early warning indications that are observed significantly before internal credit rating changes are of high importance for banks (e.g., Nakamura (1993)). Based on this information loan officers can either take restrictive actions or make an attempt to help the borrower through a period of financial distress. For example, banks may want to stop the relationship, to reduce or to cancel credit lines, to require additional collateral or to increase the existing credit line. In this context, we investigate by means of event study methodology and multivariate probit regressions models if and how early checking account variables indicate an abnormal pattern before default events. We also differentiate the usefulness of checking account information by borrower and default type. Furthermore, we condition on the initial and the pre-default internal credit ratings and control for bank relationship characteristics like the number of accounts, distance and duration. Our study contributes in several ways. This is the first extensive test of the usefulness of checking account information for default predictions. Thanks to our data we can analyze different types of borrowers (large firms, small businesses, and individuals; implying different lending technologies like secured vs. unsecured, transactions-based vs. relationship lending; e.g., Berger and Udell, 2006), different default events, and an institutional environment in which checking accounts are considered as the core of a bank-customer relationship. In a 3

4 related study, Mester, Nakamura, and Renault (2007) examine transactions accounts of small firms in the context of one particular lending technology (asset-based lending). For this institutional setting, they provide first empirical evidence in favor of the checking account hypothesis. Our analysis is based on a unique data set, including more than 3 million accountmonth observations, and suggests that the scope of the checking account hypothesis goes beyond small firms and also holds for different lending technologies. Moreover, we introduce new measures based on payment variables, e.g., the amplitude of monthly account balances and the credit-to-limit ratio that help to predict future borrower defaults. Our data is particularly appropriate because in Germany, the world s third largest banking sector behind the U.S. and Japan (as of year-end 2006), checking accounts are of key importance for firms and individuals: all incoming and outgoing payments are routed through this account. 1 Credit cards do not play a significant role for the payments of firms and are, relative to the U.S., less important for individuals in Germany. Eventually, small businesses typically borrow from Hausbanks that act as relationship lenders (e.g., Elsas and Krahnen (1998), Elsas (2005)). The remainder of the paper is organized as follows. Section 2 reviews the related literature and Section 3 describes the data. In Section 4 we explain the methodology and report main empirical results. Section 5 provides robustness tests and Section 6 concludes. 2. Review of related literature This study builds upon three strands of literature. First, we briefly review empirical studies which include proxy variables to measure the closeness of bank relationships. Second, we focus on the relatively scarce literature that investigates the role of checking accounts as a source of information for banks. Third, we summarize insights about the importance of 1 Strikingly, 99% of all non-cash transactions of firms and individuals in Germany (number and value of transactions) are processed through checking accounts (by means of wire transfers, direct debits and debit cards; Deutsche Bundesbank, 2008). For comparison, credit cards account for 14% of the total number (25% of the total value) of all payment card transactions in the year

5 checking account data for banks internal rating systems from the applied credit risk and the banking regulation literature. The first category of studies is based on the idea that banks can gather information if the borrower used or has been using different, i.e. non-credit related, financial services. 2 In other words, there may be an informational spillover from other financial services to the lending business. Examples for these other financial services are checking accounts, payment services, savings and money market accounts, brokerage, and underwriting activities (e.g., Black, 1975; Fama, 1985; Vale, 1993; Nakamura, 1993; Petersen and Rajan, 1994; Blackwell and Winters, 1997; Cole, 1998; Chakravarty and Scott, 1999; Petersen and Rajan, 2002; Cole, Goldberg, and White, 2004; Elsas, 2005; Berger, Miller, Petersen, Rajan, and Stein, 2005; Bharath, Dahiya, Saunders, and Srinivasan, 2007). These analyses explicitly hint at the specific source of a potential informational advantage of banks and include dummy variable in empirical models to measure its impact. However, existing studies provide mixed evidence on the question whether checking accounts are useful for banks to gather information. In addition, more recent research indicates that borrowers having checking accounts at their lenders exhibit a smaller distance to their bank and communicate in a more personal way with their bankers (Petersen and Rajan, 2002). Moreover, firms with checking accounts at small banks benefit from higher credit availability (Cole, Goldberg, and White, 2004) and longer loan maturities (Kirschenmann and Norden, 2008). Elsas (2005) shows that payments and informational financial services have a positive impact on the probability of being a housebank. Finally, Berger, Miller, Petersen, Rajan, and Stein (2005) find a negative impact of checking account dummies on bank size, the physical distance between the bank and firm, the probability of impersonal communication, and a significantly positive impact on the probability of having an exclusive lender and on the duration of the bank relationship. As 2 Subsequently, we focus on information synergies. Interestingly, there is also evidence for liquidity and cost synergies (e.g., Berlin and Mester, 1999; Kashyap, Rajan, and Stein, 2002), providing rationales for the simultaneous supply of deposit taking and lending by banks. 5

6 mentioned above, these studies are restricted in the sense that they include dummy variables that indicate the existence of checking accounts without explaining how banks can get an informational advantage. This paper looks inside the checking accounts to detect a link between the payment behavior of bank borrowers and their credit risk. The second strand of literature goes one step further in identifying the mechanism which may provide banks with useful credit-relevant information. Black (1975), stating that if the individual routes most of his receipts and payments through his loan account, they can serve as a continuing source of credit information, and Fama (1985) have inspired Nakamura (1993) to propose the checking account hypothesis. The latter states that checking accounts may be useful for banks to monitor small borrowers. Concurrently, Vale (1993) proposes a theoretical model to study the complementary role of deposits as a source of funding and private information in lending. Solving this model leads to two empirical implications. First, borrowers that have been depositors receive better loan terms than others. Second, exclusive borrowers are treated more favorably because they do not hide information from their banks. Mester, Nakamura, and Renault (2007) provide first empirical evidence in favor of the checking account hypothesis. They analyze the usefulness of checking account information for a small sample of firms borrowing from a Canadian bank during the period The most important difference to our study is that Mester, Nakamura, and Renault (2007) consider one particular lending technology: the case of asset-based lending (secured lending with receivables and inventories as collateral). The analysis is based on annual data from 100 small firms, hereof 50 in financial distress (loan specifics, ratings, dates of reviews and completion of reviews, financial statements, information on the exclusivity of the relationship), and monthly data (value that the bank assigns to receivables and inventories on the basis of information provided by the borrower, end-of-month balance in the borrower s checking account as well as minimum, average, and maximum monthly balances). The study has the following key results. Monthly changes in receivables can be retraced in checking 6

7 accounts if the borrower has an exclusive bank relationship. Moreover, borrowings that exceed inventory and receivables (which serve as a collateral for the credit line) can predict rating downgrades and loan write downs. In addition, the analyzed bank intensifies its monitoring as the credit quality decreases, i.e. loan reviews become lengthier and more frequent. Sufi (2007) takes a corporate finance perspective, analyzing credit lines by means of 10-K SEC filings from 4,604 U.S. firms during the period The main finding is that firms which have access to credit lines from banks have to maintain a high cash flow level to be able to comply with covenants. In contrast, firms with low cash flows or covenant violations due to declines in cash flows face a more difficult access to credit lines. Overall, these results suggest that the lack of access to a credit line is a robust measure to describe financial constraints of firms. The third strand of literature investigates, from an applied risk management or regulatory perspective, whether checking account variables (e.g. account balances, turnover measures for debits and credits, days of unauthorized overdrafts) can serve as an early warning indicator, whether it can help to improve banks internal rating systems, and how banks can estimate exposure at default (EAD) and credit conversion factors (e.g., Eisfeld, 1935; Apilado, Warner, and Dauten, 1974; von Stein, 1983; Hackl and Schmoll, 1990; Schlüter, 2005; Jiménez, Lopez, and Saurina, 2007). Note that these studies mainly examine methodological issues and statistical measurement techniques while the first and second strand of literature focus on the economic importance of checking account information for banks. Apilado, Warner, and Dauten (1974) show that the possession of a checking account is an important indicator of the creditworthiness of individual borrowers. Von Stein (1983) provides first quantitative evidence that checking account data allows to discriminate between defaulting and nondefaulting firms already two years before default. Subsequently, Hackl and Schmoll (1990) are able to confirm these findings. Moreover, Jiménez, Lopez, and Saurina (2007) analyze data from the Spanish Credit Registry during the period to provide evidence on 7

8 corporate credit line utilization and its implications for exposure at default (EAD) estimations. They show that the risk profile of the borrower, characteristics of the bank, and the business cycle have a significant impact on credit line utilization. 3. The data This study is based on a unique data set with monthly observations from checking accounts provided by a German universal bank. The bank is among the largest 5% by total assets in the category of comparable banks, as defined by the Deutsche Bundesbank. The sample period covers five calendar years, starting from January 2002 until December Initially, we start with a data set including more than 3.7 millions of account-month observations and apply the following selection and adjustment rules. First, we exclude all borrowers that are financial institutions, public entities (state or municipality-owned), and certain legal advisers (lawyers, solicitors, notaries) because we want to analyze checking accounts of commercial and individual borrowers. Legal professions are excluded because notaries typically hold a large number of special purpose transaction accounts to enable their clients to meet obligations from sales contracts (for example real-estate transactions). Second, we consider all customers borrowing from the bank during the period to ensure that our data set is free of a survivorship bias that may arise from bank switching and default events. 3 Third, we differentiate borrowers by (i) legal form (categories: GROUP = 1 for companies like corporations and partnerships, GROUP = 2 for small businesses and liberal professions, and GROUP = 3 for individuals) and (ii) size measured by the mean credit payments per month (categories: SIZE = 1 for large borrowers (>95% quantile: 12,835 EUR), SIZE = 2 for mid-sized borrowers (75% to 95% quantile: between 2,979 EUR and 12,835 3 The data selection is not based on a particular point in time but the entire sample period. The sample includes bank borrowers who started borrowing before 2002 but which are no longer in the sample in subsequent years because they switched the bank or because they went bankrupt. Moreover, the sample also includes borrowers who have started borrowing from the bank during For example, there are 58,790 (51,069) accounts with a time series of at least 36 (48) months. We have data on exactly 60 months for 6,235 accounts. 8

9 EUR), and SIZE = 3 for small borrowers (0% to 75% quantile: between 0 EUR and 2,979 EUR). On the one hand, differentiating by legal form has the clear advantage that this classification is not based on data from checking accounts, i.e. it can be seen as exogenous relative to the payment behavior. On the other hand, differentiating borrowers by their mean credit payments may represent a more accurate proxy for size than the legal form because many small firms in Germany are corporations and but there are also some large businesses that are no corporations. However, the size proxy may be biased if a firm has multiple bank relationships and most of the credit payments are not observed. As both arguments are valid, we differentiate by GROUP and SIZE. 4 Fourth, if borrowers change their legal form (for example, an individual starts running small business or a partnership is transformed into a corporation), we assign the legal form that is observed in most of the months to the account and the borrower. Finally, we do not have information about a borrower s number of bank relationships. We believe that this is not a problem because it creates a conservative bias in our study, making it more difficult to provide evidence on the usefulness of checking account information for credit risk monitoring. If we assume that some borrowers have multiple bank relationships and the analysis of data from one bank reveals that checking account information is useful, the corresponding results should be even stronger for a homogeneous sample including only borrowers with single bank relationships. Moreover, given the fact that most individuals and small businesses in Germany typically have only one or two checking accounts, the dilution effect should not be very strong. Summary statistics are reported in Table 1. Insert Table 1 here 4 The variable GROUP (based on the legal form) and SIZE (based on mean credit payments) exhibit a rank correlation of 0.28 (p<0.0001), indicating that both criteria are significantly positively related but not perfectly correlated. Since the SIZE is based on the time-series mean of an account this classifier is time-invariant. 9

10 Panel A describes the structure of our final data set. As can be seen from the rightmost column, it includes 3,271,879 account-month observations from 86,945 accounts (67,215 borrowers). Columns 2-4 in the upper part of Panel A differentiate the composition of the sample by legal form. 3.79% of all account-month observations come from companies, 10.28% from small businesses, and 85.93% from individuals. Given this large share of individuals, it is not surprising that most of the borrowers (78%) have just one checking account while companies and small businesses are typically among those who have multiple accounts. Note that we will study the robustness of the results with respect to the number of accounts later in more detail. Moreover, an important feature of our data set is that we have a time series of months for 77% of all checking accounts, allowing not only crosssectional but also time-series analysis. Panel B summarizes the main checking account and bank-relationship variables for the entire sample and by borrower type at the account level. Note that all checking account variables are reported in thousands of Euro. For the entire sample, the mean minimum balance per month (LOW) is and the corresponding maximum (HIGH) is Mean cumulative debits (DEBIT) and credits (CREDIT) are 8.04 and 8.22 respectively. The mean balance (MID) is 1.99 and the average credit line (LIMIT) amounts to (LIMIT is defined as a negative number). Comparing variables across the three categories of borrower types (by GROUP which based on the legal form) shows that companies have the largest mean credit payments (CREDIT=119.15) and individuals the smallest (CREDIT=3.34). The same is true for LOW, HIGH, DEBIT and LIMIT, indicating that the classification based on GROUP is reasonable. In the entire sample, the mean duration of the bank-borrower relationship DUR is 8.06 years 5 and the mean distance between the domicile of a borrower and the bank s head office 5 Related studies report similar durations, for example 7.9 years (Degryse and Ongena (2005)) and 5.1 years (Ongena and Smith, 2001). 10

11 DIST is 7.21 kilometers. 6 The average internal credit rating RAT is 2.87 (on a 6-grade scale, with 1 being the best grade, grades 5 and 6 included defaulted borrowers) and most of the borrowers have a rating of 2 or 3. The bank has different internal credit rating systems for firms and individuals and systematically considers the rating for all loan approval decisions, loan pricing and loan loss provisioning (e.g., Machauer and Weber, 1998; Treacy and Carey, 2000; Grunert, Norden, and Weber, 2005). Internal credit ratings for firms are based on hard facts (analysis of financial statements) and soft facts (market position, management quality, etc.) while ratings for individuals are based on variables like age, marital status, home status, income, etc. Most important, information from checking accounts is not included in the bank s internal credit rating system and it is not used in any other systematic way. This fact rules out that borrowers strategically adjust their payment behavior to hide information from the bank. In addition, borrowers which are close to financial distress have the strongest incentives to hide information but they have typically the lowest financial flexibility to do so. Panel C presents the number of rating changes and default events. During the sample period, we observe 12,803 changes of the internal credit ratings. Although we know the exact dates, we assign a rating change to the corresponding month since all other variables are measured at a monthly frequency. Differentiating by direction of the rating change leads to 5,515 rating upgrades and 7,288 rating downgrades. We observe 1,009 default events, i.e. these borrowers were downgraded to grade 5 (first specific loan loss provision, 90 days past due on any obligation, restructuring, utilization of collateral, attachment proceedings) or grade 6 (bankruptcy filing). The definition of default refers to the borrower level and is common practice at banks and meets the international regulatory requirements (Basel Committee on Banking Supervision, 2006). The type of default is distributed almost uniformly with The variable DIST reflects the aerial distance between the borrowers and the bank, based on the first three digits of the ZIP code. Other studies report 8 miles (Agarwal and Hauswald, 2007; for relationship borrowers), 9 miles (Petersen and Rajan, 2002), and 4.3 kilometers (Degryse and Ongena, 2005). Our results are not much affected if we compute the distance between the borrower and the nearest bank branch because all branches of the bank are relatively close to the bank s head office which is located in a metropolitan area. 11

12 downgrades to rating 5 and 468 downgrades to rating 6. In addition, there are 168 companies, 211 small businesses, and 630 individuals that default during our sample period. Finally, note that defaults to grade 5 are under some control of the bank management 7 whereas defaults to grade 6 (bankruptcy filing) can be seen as exogenous events which may have implications for the usefulness of checking account information. 4. Empirical analysis 4.1. Methodology and univariate results We consider two ways of measuring whether checking account information is useful for assessing the credit risk of bank borrowers. First, we apply event study methodology which has been extensively used to analyze anticipation and announcement effects in financial economics. This approach enjoys the advantage that the results can be easily visualized. In addition, we estimate multivariate probit regression models in calendar time (for different time horizons) to investigate which factors influence the probability of going to default in month t+1. In the event study, we proceed as follows. First, we identify and mark default events in calendar time. Second, we transform calendar time into event time at a monthly frequency with an event time window of 48 months [event time = -36, -35,, 0, 1,, 12]. During the period we observe 1,009 default events and assign these incidents to event time 0. Third, we calculate different checking account variables of borrowers that default at event time 0 for each month in the event time window. For comparison purposes, we calculate the same variables at the same month for all borrowers in our sample. The latter serve as a benchmark in order to identify abnormal patterns of checking account variables of defaulters. This approach has the clear advantage that a bank can ex ante compare each borrower with an 7 Note that the loan loss provisioning follows strict regulations and the actual specific loan loss provisions made by the bank are checked by the auditors, tax authorities, and federal banking supervisors. 12

13 appropriate benchmark (average of all borrowers, industry-specific, size-based benchmarks). Note that this methodology is similar to calculating abnormal stock returns (with a stock index return, an average portfolio return or a market model return as a benchmark). Taking non-defaulting borrowers as a benchmark suffers from the problem that a bank can identify this group only from an ex post perspective, i.e. it does not help banks obtaining ex ante warning indications. 8 Subsequently, we report results which are based on the median variables for defaulters and all borrowers for the entire sample. 9 In a first step, we examine the credit line usage and the cumulative number of credit limit violations. We measure credit line usage (in %) by the variable USAGE_LOW = LOW / LIMIT and the variable USAGE = MID / LIMIT. A credit limit violation occurs if the minimum balance in a month falls short of the credit limit (LOW < LIMIT). The cumulative number of credit limit violations CUMVIOL is defined as the sum of all monthly limit violations up to a particular month. Figure 1 displays the findings for the medians of these variables. Insert Figure 1 here It turns out from Figure 1a that the median credit line usage of defaulters is very different to those of non-defaulting borrowers. While USAGE_LOW (USAGE) of defaulters start above 60% (20%) 36 months prior to default, the corresponding values of non-defaulters amount to roughly 0% (-30%). As explained above a negative sign for credit line usage indicate that there is no usage at all, i.e. the checking account has a credit balance. If we come closer to the default at event time 0, the value of USAGE_LOW (USAGE) for defaulters runs 8 We also calculated the corresponding checking account variables for (i) all non-defaulters and (ii) all borrowers (without those that are already in default). The results are almost identical to the reported ones since the impact of the defaulters on the average variables of the entire sample is very small. 9 Differentiated results by borrower type (legal form and size) are reported in Section

14 up to almost 100% (80%), indicating a systematic increase of credit line usage. This observation indicates that borrowers which subsequently default exhibit a gradually increasing need of liquidity which is, what we will show later, due to a decrease in credit payments (e.g., sales, salary). Most of the run up of USAGE_LOW occurs during the event time interval [-36, -16]. In addition, USAGE displays a sharp increase during the nine months prior to default. Most important, during negative event time the credit line usage (USAGE_LOW and USAGE) of all borrowers remains relatively stable around 0% (-30%). It can be seen that there is no systematic increase or decrease in the benchmark. This finding is consistent with the results for Spanish firms by Jiménez, Lopez, and Saurina (2007) who detect a major increase in credit line usage during the 12 months prior to default. Moreover, from Figure 1b it can be seen that the cumulative number of credit limit violations for defaulters differ considerably from those of all borrowers. While the latter remains relative flat, the former increases continuously from a median of 5 at event time -36 to a median of 20 at event time 0. Note that the slope of the curve for defaulters becomes considerably steeper around event time -18. This relatively strong result is in line with the findings on Canadian firms from Mester, Nakamura, and Renault (2007). Overall, we conclude that defaulters exhibit a significantly abnormal increase in credit line usage and in the number of limit violations before default. In a next step, we analyze another potentially useful checking account variable that has not been considered in related studies so far. In order to asses the variation of balances within a month, we calculate the absolute account amplitude AMPLI, defined as AMPLI = HIGH LOW. This variable takes always positive values regardless of the sign of the minimum LOW and maximum HIGH. Our hypothesis is that borrowers that go bankrupt exhibit a systematic decrease in the absolute account amplitude. This may be the case for the following to reasons. First, firms as well as individuals which subsequently default are expected to face a decline in credit payments. In other words, firms get into financial distress mainly if sales decrease (i.e. 14

15 credits decrease) while debits remain relatively constant. The same reasoning holds for individuals that become unemployed, facing no or a reduced transfer income. Second, given a decrease in credit payments borrowers become increasingly financially constrained. Consequently, from a certain level debits have to decrease as well in order avoid overdrafts, encumbrance etc. Figure 2 depicts the event study results for the median of AMPLI. Insert Figure 2 here Interestingly, the median amplitude AMPLI of defaulters and all borrowers are relatively similar and in a range of 1,000 to 2,000 EUR during the event time interval -36 to -6. The fact that AMPLI for defaulters is above the other curve during event time -36 to -30 is due to a relatively small number of observations in the first months. 10 More important, from event time -5, we observe that the AMPLI of defaulters drops considerably from 1,000 EUR to less than 200 EUR while the corresponding value for all borrowers does not change at all. Based on these results, we conclude that the account amplitude provides a warning indication roughly 5 months prior to default. In addition, we examine not only account balances but also turnover information reflected by the cumulative monthly credit and debit payments (CREDIT, DEBIT). We divide CREDIT and DEBIT respectively by the credit line LIMIT to evaluate the relative evolution of payments at defaulters to those at all borrowers. 11 Note that no other study has analyzed payment variables like CLR and DLR before. Results for the medians of both variables are presented in Figure The graph for defaulters is based on 310 (574) observations at event time -36 (-24) and on average, during the entire event time window on 771 observations. 11 The inverse of CLR and DLR can be interpreted as limit-credit duration (how many months does it take that cumulative credit payments can pay back the limit?) and limit-debit duration (how many months does it take that cumulative debit payments reach the limit?). 15

16 Insert Figure 3 here Figure 3 indicates that both debits and credits (relative to the credit line) of defaulters are more or less stable in the range 60% to 80% during event time -36 to -18 and they drop sharply afterwards. Most of this sudden decrease occurs during event time -18 to -12. Obviously, cumulative credits and debits are highly correlated and move relatively in tandem as hypothesized above. For comparison, CLR and DLR of all borrowers do not change very much, remaining above 70% all the time. Given these findings, we conclude that the evolution of credit payments indicates an abnormal pattern roughly months prior to the actual default event Multivariate regression analysis We now turn to a multivariate econometric analysis by means of probit regression models. All regression models are estimated in calendar time in order to take an ex ante perspective. Subsequently, we study which factors at calendar time t and t-12 influence the probability of default in calendar time t+1. Default is indicated by a dummy variable which takes a value of one if the borrower exhibits a jump to the default grades (rating 5 or 6) in t+1. Explanatory variables are the internal credit rating (RAT), and changes of the internal rating (ΔRAT), the credit line usage (ΔUSAGE), the absolute checking account amplitude (ΔAMPLI), the creditto-limit ratio (ΔCLR) and the cumulative number of limit violations (ΔCUMVIOL) during the period [t, t-12] and the period [t-12, t-24]. The number of observations decreases since we cannot include accounts (i) with a history of less than 13 or less than 25 months respectively and (ii) without a credit line. However, in Section 5.2 we summarize results from an alternative model that is based on considerably more observations. We also include the percentage change of the credit line (ΔLIMIT) as a control variable. This is important because 16

17 the checking account variables (except AMPLI) are based on the payment behavior and the credit limit. However, the bank can change the credit limit to manage the loan exposure. For example, USAGE may increase if (i) the account balance steadily declines while LIMIT remains unchanged (i.e. the bank does not manage the loan exposure) or (ii) the account balance remains stable while LIMIT has been reduced by the bank. We expect a negative sign for the coefficient of limit changes because limit reductions can be one response of the bank with regard to borrowers facing financial problems. All variable changes are calculated over twelve months because within this period at least one internal rating review has to be carried out. 12 Specifically, we estimate the following models to predict borrower default in month t+1: Model I including only the level of the internal rating 13 in month t, Model II including only changes of checking account variables 14 over the last 12 months, Model III including the level of the internal rating in t and changes of checking account variables over the last 12 months, and Model IV including the changes of the internal rating and changes of checking account variables over the last 12 months (Table 2, Panel A). Furthermore, we also consider different prediction horizons (t-12 in Panel B; t and t-12 in Panel C). Note that the internal credit rating (and/or rating changes) can be seen as a benchmark against which we test the usefulness of checking account information. 15 Results are reported in Table 2. Insert Table 2 here 12 This approach is conservative because we do not consider monthly variable changes which are more likely to occur at the checking accounts than in case of the internal ratings. Nonetheless, we have also estimated the model with changes of all variables over consecutive 3 and 6-months intervals and obtain very similar results. 13 We include the ordinal rating variable (scale from 1 to 6). Alternatively, if we include dummy variables with either grade 1 or grade 3 as reference category all of the subsequently reported results remain unchanged. 14 We believe it is reasonable to use dynamic information (changes) rather than static information (levels) to fully exploit the usefulness of checking accounts. Nonetheless, we have also estimated models including levels of checking account variables and obtain basically similar results. 15 In Section 4.6 we describe results from a two-stage regression model that confirms our basic findings. 17

18 The two key findings are that checking account variables have a significant impact on the probability of default in addition to the internal credit rating and that their inclusion increases the goodness-of-fit of the default prediction models substantially. It can be seen from Panel A that, as expected, the internal credit rating is significantly positively associated with future borrower defaults (Model I). 16 Most important, Models II, III, IV indicate that all checking account variables are correctly signed and highly significant. Consistent with previous findings, the change of the credit line usage and the cumulative number of limit violations display a positive coefficient and the change of the absolute amplitude and the credit-to-limit ratio a negative coefficient. As expected, the control variable ΔLIMIT is statistically significant and negative. Furthermore, comparing the goodness-of-fit of the models (based on the McFadden-R 2 adjusted for the number of regressors) reveals that using checking account information alone leads already to a better fit than the rating level alone. Strikingly, combining both sources of information more than doubles the McFadden-R 2 (Model I: vs. Model III: or Model IV: 0.070). From an economic perspective, this finding suggest that banks can improve their monitoring and default predictions by including checking account and payment variables in their borrower monitoring systems. The corresponding regressions models in Panel B include the explanatory variables lagged by 12 months, i.e. the rating level at month t-12 and changes of variables during the period [t-12, t-24]. The regression models displayed in Panel C include information from the year before default and two years before default. Note that we cannot estimate Model I including rating levels in t and t-12 since these are highly correlated, leading to problems of multicollinearity. Essentially, results from Panel B and C are consistent with the hypothesis that checking account information is useful to predict borrower defaults. One additional noteworthy finding is that the account amplitude (AMPLI) is a good short-run default predictor (see Figure 2). The 16 Note that positive rating changes indicate a credit quality deterioration (for example, from rating grade 2 at time t-12 to rating grade 4 at time t) since higher numbers correspond to rating grades of a higher default risk. 18

19 change of this variable is statistically significant and correctly signed in the year before default but it becomes insignificant over the period t-12 to t-24 (see Panel B and C). In summary, estimation results from various probit regression models confirm the univariate results from the event study. In addition, the multivariate analysis shows that checking account information remains useful if we control for levels and changes of internal ratings as well as changes of credit lines Results by borrower type and size We now continue the analysis by examining whether our previous results hold for different types of borrowers. As discussed beforehand we differentiate between both the legal form GROUP (companies, small businesses, and individuals) and SIZE (large, mid, and small). The complexity of payments at the checking accounts, the likelihood of having multiple banking relationships, and the mechanism of default differ considerably across these groups. Moreover, banks typically apply different lending technologies for firms and individuals (e.g., Berger and Udell, 2006). Also note that credit lines granted to firms are frequently secured by means of a collateral pool (and the value of this pool is independent of the actual credit line usage) while credit lines associated with checking accounts of individuals are typically unsecured. The size of our data set and, more important, these institutional differences across borrower groups allow us to carry out an extensive test of the usefulness of checking account information while the evidence provided by Mester, Nakamura, and Renault (2007) is related to the case of asset-based lending (where the extension of credit is primarily based on the value of collateral rather than the overall credit quality of the borrower; e.g. Berger and Udell, 2006). In the following event study we use GROUP- or SIZE-specific benchmarks in this section. For example, a checking account variable of defaulters in SIZE=1 (or GROUP=1) is compared with the median of all borrowers from the same category, i.e. with SIZE=1 19

20 (GROUP=1). This analysis is not only interesting for study differences across groups but it is also more accurate than the analysis from the previous section because the benchmarks are group-specific. Figure 4 displays the event study results for the absolute checking account amplitude AMPLI which can be interpreted as an indicator for account balance variation within a month. 17 Insert Figure 4 here Essentially, findings from the previous section are confirmed. In all sub-groups we find that the absolute amplitude of defaulters decreases considerably some time before default while the one of all borrowers does not change very much. However, there are important differences within the categories of GROUP and SIZE. Differentiating by legal forms based on GROUP (Figures 4a, b, c) reveals that the warning sign by AMPLI occurs 12 to 6 months before default at small businesses, 6 months before default at individuals and only slightly before default at companies. A differentiation by SIZE (Figures 4d, e, f) leads to very similar results. However, for small borrowers (essentially individuals) a clear warning indication is observed 12 months prior to default. In addition, it turns out that there are no clear warning indications for companies and large borrowers. In summary, we conclude that checking account information from small businesses and individuals (mid-sized and small borrowers) provide banks with warning signs about the credit quality 6-12 months prior to a subsequent default. In contrast, checking account information does not significantly help monitoring companies (corporations, partnerships) and large borrowers (with mean monthly cumulative credit payments above 12,835 EUR). 17 We have repeated the univariate analysis with the variables USAGE, CLR and DLR and obtain qualitatively similar results. 20

21 In a next step, we re-estimate the probit regression models from the previous section for the categories of GROUP and SIZE. Table 3 summarizes the estimation results. We only report results for the regression model corresponding to Model IV (Table 2, Panel C) to conserve space; results for the other models lead to very similar outcomes. Insert Table 3 here Table 3 provides very interesting insights on the usefulness of checking information across types of borrowers. Panel A presents the results differentiated by GROUP. The change of the internal credit rating, the credit line usage and the cumulative number of violations display a significantly positive coefficient for companies. For small businesses, only ΔCUMVIOL at t and t-12 and the internal credit rating change at t-12 have a significantly positive influence on the probability of default. Interestingly, for individuals we find that all contemporaneous checking account variables are correctly signed and highly significant. In addition, in contrast to commercial borrowers there is no impact of the internal rating change at t-12. Panel B distinguishes by size categories. Basically, the results shown in Panel B are even clearer than in Panel A. The findings for large and mid-sized borrowers are very similar, i.e. the rating change in t and t-12 and the change of the cumulative number of violations display significant coefficients (and ΔUSAGE at t and t-12 for mid-sized borrowers). In opposite, internal credit rating changes have no impact on the probability of default for small borrowers while all checking account variables are signed as expected and highly significant. It is noteworthy that the change of the cumulative number of limit violations (ΔCUMVIOL) performs especially well. The contemporaneous change of this variable exhibits a significantly positive coefficient for all GROUP- and SIZE-categories and the lag is significant for small businesses, individuals and small borrowers. Finally, as expected the contemporaneous change of ΔLIMIT is significantly negative in for all categories. Our results 21

22 on the checking account variables do not change if we exclude ΔLIMIT from the regression models. Given these differences across borrower types we now investigate what banks can gain in economic terms by using checking account information for borrower monitoring. In particular, we want to test whether the relative gain of using checking account information is highest for monitoring small business and individuals, as suggested by the previous analysis. More specifically, we examine how the goodness-of-fit (adjusted McFadden-R 2 ) of default prediction models changes if a bank includes checking account in addition to internal credit ratings. Table 4 summarizes the corresponding findings. Insert Table 4 here This analysis reveals a very clear pattern. Consistent with Table 3 it turns out that banks can gain most if they use checking account information for monitoring small businesses and individuals. Comparing the values of the adjusted McFadden-R 2 displayed in Table 4 indicates that there is even a considerable increase in the case of companies but the relative gain is largest for small businesses and individuals. This finding is consistent with the fact that default prediction models including the internal credit rating only (or rating changes) perform best for companies (large borrowers). The economic explanation is that input factors of the credit rating for companies (accounting information, private soft information of the bank) tend to be more sensitive to a deterioration of a borrower s credit quality than information from checking accounts (that may also be diluted across different bank relationships) while the opposite holds for small businesses and individuals. For these borrowers the input factors of the credit rating seem to be sticky while checking account information reflects changes of a borrower s credit quality in a relatively timely manner. 22

23 Summarizing, we find that checking account information is particular useful for monitoring small businesses and individuals and the relative improvement of default prediction models after including checking account information is largest for these borrowers Results conditional on internal credit ratings Subsequently, we investigate whether the usefulness of checking account information depends on (i) the credit rating in the default status, i.e. the type of default event, (ii) the rating at the beginning of the checking account time series, and (iii) the rating one month prior to default. Our expectation is that checking account information is more useful for (i) defaults to rating grade 6 (bankruptcy filing) in comparison to defaults to rating grade 5, (ii) the better the rating at the beginning, and (iii) the better the rating before default. We start the analysis with a differentiation by default event types. As mentioned above default to rating grade 5 is under some control of the bank management while a default to rating grade 6 can be regarded as exogenous and has more severe consequences for the bank. Figure 5 shows event study results for the credit line usage behavior (USAGE) if we differentiate the 1,009 default events by default type (541 defaults to rating grade 5 and 468 defaults to rating grade 6). Insert Figure 5 here During the event time -36 to -12 months the credit line usage USAGE increases in a very similar way that significantly differs from that of all borrowers. It is noteworthy that the credit line usage of borrowers with subsequent defaults to rating grade 6 is above the other curve most of the time. Interestingly, it increases sharply from 50% to almost 100% during the 9 months prior to default while the corresponding variable for defaults to rating grade 5 even decreases during the event time interval -4 to 0. In summary, we find that credit line usage 23